On 07.11.2019 12:30, 曾文旌(义从) wrote:

May be the assumption is that all indexes has to be created before GTT start to 
be used.
Yes, Currently, GTT's index is only supported and created in an empty table 
state, and other sessions are not using it.
There has two improvements pointer:
1 Index can create on GTT(A) when the GTT(A)  in the current session is not 
empty, requiring the GTT table to be empty in the other session.
Index_build needs to be done in the current session just like a normal table. 
This improvement is relatively easy.

2 Index can create on GTT(A)  when more than one session are using this GTT(A).
Because when I'm done creating an index of the GTT in this session and setting 
it to be an valid index, it's not true for the GTT in other sessions.
Indexes on gtt in other sessions require "rebuild_index" before using it.
I don't have a better solution right now, maybe you have some suggestions.
It is possible to create index on demand:

Buffer
_bt_getbuf(Relation rel, BlockNumber blkno, int access)
{
    Buffer        buf;

    if (blkno != P_NEW)
    {
        /* Read an existing block of the relation */
        buf = ReadBuffer(rel, blkno);
        /* Session temporary relation may be not yet initialized for this backend. */         if (blkno == BTREE_METAPAGE && GlobalTempRelationPageIsNotInitialized(rel, BufferGetPage(buf)))
        {
            Relation heap = RelationIdGetRelation(rel->rd_index->indrelid);
            ReleaseBuffer(buf);
            DropRelFileNodeLocalBuffers(rel->rd_node, MAIN_FORKNUM, blkno);
            btbuild(heap, rel, BuildIndexInfo(rel));
            RelationClose(heap);
            buf = ReadBuffer(rel, blkno);
            LockBuffer(buf, access);
        }
        else
        {
            LockBuffer(buf, access);
            _bt_checkpage(rel, buf);
        }
    }
    ...


This code initializes B-Tree and load data in it when GTT index is access and is not initialized yet.
It looks a little bit hacker but it works.

I also wonder why you are keeping information about GTT in shared memory. Looks like the only information we really need to share is table's metadata. But it is already shared though catalog. All other GTT related information is private to backend so I do not see reasons to place it in shared memory.

--
Konstantin Knizhnik
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company



Reply via email to